Research On Continuous Tracking Of Human Target In Distributed Cameras

According to the problem of human moving target detecting and persistent target tracking in thedistributed cameras, this dissertation focuses on human target detecting and tracking in video imagesand target handoff in distributed cameras. Through detecting and judging human target in each videoimage of the distributed cameras, persistent target tracking in the large-scale scenes is achieved.The main contents of this dissertation are as follows:1、The target detection and tracking in a single camera is studied. At first, the area of movingtarget is detected by subtraction from the background, and then each human region is extracted bydifferent characteristics of the human model. Template matching tracking algorithm and targettracking algorithm based on Mean Shift and Kalman filter are emphatically studied. The results showthat target tracking algorithm based on Mean Shift and Kalman filter is characteristic of betterreal-time and greater robustness. And this method can solve the target occlusion problem. In a word,this method can basically meet the requirement of real-time for human target tracking.2、Target handoff in the distributed cameras. Target handoff algorithms based on the projectiveinvariants and homographic transformation are discussed. There are two problems to be handled in theprojective invariants algorithm. One is inaccuracy of selecting points artificially and the other ismultiple targets at close range will go wrong. In order to improve the algorithm, Scale-invariantfeatures transform (SIFT) algorithm and strategy of main hue analyzing are introduced. In order toeliminate mismatching points, RANSAC method is presented and then the best homographic matrix isgot. The results show that improved scheme improves the accuracy of the target handoff algorithm.3、Human target continuous tracking in the distributed cameras. When the human target moves ina single camera, algorithm based on Mean Shift and Kalman filter is used to detect the target. Whenthe human target enters into the overlapping area of the cameras, the position of target is identifiedusing homographic transformation algorithm and then corresponding video image is selected todetecting the target continually. The experiment results show that this method is validity.